Moonshine

Moonshine is a family of speech-to-text models optimized for fast and accurate automatic speech recognition (ASR) on resource-constrained devices. It is well-suited to real-time, on-device applications like live transcription and voice command recognition. English Moonshine obtains word-error rates (WER) better than similarly-sized Tiny and Base Whisper on the OpenASR leaderboard, and non-English Moonshine variants outperform Whisper Small and Medium, which are 9x and 28x larger, respectively.

Moonshine processes audio segments between 5x-15x faster than Whisper while maintaining the same (or significantly better!) WER/CER. This is because its compute requirements scale with the length of input audio. Shorter input audio is processed faster, unlike Whisper models that process everything as 30-second chunks.

Unquantized Base is 62M parameters (or 400MB), while Tiny is 27M parameters (around 190MB).